TypeSafe · TypeSafe API
Jev 1.13 by TypeSafe scores 79.3 on the System One Index. Calibration error 5.8%. An earlier run, kept for reference.
| Dataset | Accuracy | Rank | Best on the board | Calibration error |
|---|---|---|---|---|
| AG News English | 87.5 | 92.5 | 9.2% | |
| SST-2 English | 94.5 | 97.0 | 6.5% | |
| SST-5 English | 63.5 | 62.5 | 18.1% | |
| DAIR Emotion English | 56.5 | 91.5 | 30.0% | |
| Banking77 English | 79.5 | 98.5 | 9.8% | |
| BoolQ English | 89.5 | 92.5 | 3.7% | |
| XNLI English English | 88.0 | 93.3 | 7.2% | |
| Typed decisions English | 73.1 | 79.0 | 4.3% | |
| XNLI French French | 76.7 | 90.7 | 12.2% | |
| XNLI Arabic Arabic | 69.3 | 89.3 | 14.3% | |
| MASSIVE scenario Arabic | 63.5 | 89.5 | 20.4% | |
| MASSIVE intent Arabic | 83.0 | 91.5 | 5.2% | |
| OSACT4 offensive Arabic | 89.5 | 95.5 | 5.1% | |
| Tweet sentiment Arabic | 76.0 | 80.0 | 9.8% | |
| MIRACL rerank Arabic | 87.5 | 90.5 | 7.2% | |
| MIRACL pairwise Arabic | 90.5 | 91.0 | 1.7% |
jev-1.13.0Same model as the OpenRouter run, measured two weeks earlier. Left out of the charts. Its latency is from the original run, on a different machine.
Ranks count the models charted on the leaderboard. Every model gets the same cases and the same request, under the benchmark rules.